Topic recommendation method, topic recommendation device, storage medium and electronic equipment

Abstract

The embodiment of the invention relates to a topic recommendation method, a topic recommendation device, a storage medium and electronic equipment, and relates to the technical field of artificial intelligence. The topic recommendation method comprises the steps of determining a multi-modal user portrait according to text information associated with a user, image information associated with the user and behavior information associated with the user; obtaining a current dialogue context of the user, and extracting dialogue information associated with the current dialogue context; the dialogue information comprises a topic feature, a dialogue intention and a dialogue process; and inputting the multi-modal user portrait, the current dialogue context and the dialogue information into an initial topic recommendation model, and outputting a recommended topic for the user through the initial topic recommendation model. User requirements can be deeply mined, so that the recommended topic accurately meets the user requirements.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of artificial intelligence technology. More specifically, the embodiments of the present disclosure relate to a topic recommendation method, a topic recommendation device, a computer-readable storage medium, and an electronic device. Background Art

[0002] This section is intended to provide a background or context to embodiments of the present disclosure, and no description herein is admitted to be prior art by inclusion in this section.

[0003] Topic recommendation is an intelligent auxiliary tool in the chat process. During the user conversation, through intelligent algorithms or large language models, based on the current conversation content, user interests and context understanding, it provides users with appropriate new topic suggestions in real time. The purpose of this technology is to intervene at the right time to help users find new communication directions, avoid stagnation or embarrassment in the conversation, and maintain the coherence and interest of the conversation. Summary of the invention

[0004] In the related art, there are generally the following two topic recommendation schemes:

[0005] The first solution is a personalized recommendation method based on collaborative filtering. The basic process of this solution is: first, analyze the user's interest tags and the current conversation context. Secondly, use the collaborative filtering method to make recommendations by analyzing the user's historical behavior (such as likes, comments, sharing, etc.) and the similarity with other users' behaviors. Specifically, the user's interest tags (such as "travel" and "technology") can be matched with relevant topics in the topic library based on preset rules, and the context of the current conversation can be considered to recommend appropriate topics. For example, if the user is interested in "travel" and "food" is mentioned in the conversation, the system will recommend "travel food guide" as a chat topic. At the same time, based on collaborative filtering, the system will mine the preferences of users with similar interests and behaviors, and apply this information to recommend topics that users may be interested in. For example, if other users with similar behaviors to the user show a high interest in a topic, the system will also recommend these topics to the other users.

[0006] The second solution is a social topic recommendation method based on avatars. This solution obtains the basic information and avatars of both chat users, and uses the information extraction model to extract common features of the basic information to obtain the first topic recommendation result; at the same time, the avatar style extraction model is used to extract common features of the avatars to obtain the second topic recommendation result. Finally, the recommended topic is determined by combining these two recommendation results, and the relevant topic information is crawled and the recommended topic sentence is generated and recommended to the user.

[0007] However, the first solution has the following defects:

[0008] ① Limitations in rule setting: Over-reliance on manually set rules may not fully cover the diverse needs of users and complex conversation situations. The formulation of rules needs to be continuously maintained and updated to adapt to changes in user interests, otherwise it may lead to limitations and inaccuracies in recommendations.

[0009] ② Insufficient personalization: Although collaborative filtering can make recommendations based on the similarity of user behaviors, it relies on historical data and focuses mainly on the similarity of user behaviors. Therefore, it may not be able to deeply understand users’ personalized interests and emerging needs, and it has poor support for new users or cold start problems.

[0010] ③Data sparsity problem: When dealing with new topics or users, collaborative filtering may encounter data sparsity problems, that is, insufficient user behavior information, resulting in inaccurate recommendation results. Especially when there is less user behavior information, the relevance and accuracy of the recommendations will be affected.

[0011] ④ Weak context understanding ability: Rules and collaborative filtering mainly rely on static rules and historical data, and may not be able to effectively understand the complex context in the conversation or the dynamically changing user needs, resulting in the recommended topics not being close enough to the actual content of the current conversation and the user's real-time interests.

[0012] ⑤ Recommendation uniformity: Rule-based and collaborative filtering methods may not be able to discover users’ potential interests and new topic trends. The recommendation results may be too single and fail to provide sufficient diversity and innovation, thus affecting the user’s interactive experience.

[0013] The second solution has the following defects:

[0014] ① Limitations of basic information and avatars: Feature extraction based only on basic information and avatars may not fully capture the interests and preferences of users, resulting in the recommended topics being incomplete or in personalized.

[0015] ②Accuracy of crawled information: The quality and timeliness of crawled topic information directly affect the relevance and accuracy of recommended topics. There may be a risk of outdated or irrelevant information being recommended.

[0016] ③ Ignoring conversation content: This solution does not combine the user’s current conversation content in the recommendation process, and ignores the interests shown in historical conversations when extracting interests, which may result in recommended topics that do not meet the user’s latest interests and needs.

[0017] Therefore, a topic recommendation method is needed so that the recommended topics can be close to the actual needs of users.

[0018] In this context, embodiments of the present disclosure are intended to provide a topic recommendation method, a topic recommendation device, a computer-readable storage medium, and an electronic device.

[0019] According to a first aspect of the embodiments of the present disclosure, a topic recommendation method is provided, comprising:

[0020] Determine a multimodal user portrait based on the text information associated with the user, the image information associated with the user, and the behavior information associated with the user;

[0021] Acquire the current conversation context of the user, and extract conversation information associated with the current conversation context; the conversation information includes: topic features, conversation intention and conversation process;

[0022] The multimodal user portrait, the current conversation context and the conversation information are input into an initial topic recommendation model, and a recommended topic for the user is output through the initial topic recommendation model.

[0023] In an optional implementation, determining the multimodal user portrait according to the text information associated with the user, the image information associated with the user, and the behavior information associated with the user includes:

[0024] Acquire text information associated with the user, and perform feature extraction on the text information to obtain the user-associated text feature; the text information includes at least one of the following: basic information of the user, user tags, and user historical conversation data; the user-associated text feature is used to describe at least one piece of information of the user: user emotional features, user communication style, and user interest features;

[0025] Acquire image information associated with the user, and perform feature extraction on the image information to obtain features of the user associated image;

[0026] Acquire the behavior information associated with the user, and perform statistical analysis on the behavior information to obtain the behavior characteristics associated with the user;

[0027] The user-associated text features, the user-associated image features, and the user-associated behavior features are fused to obtain the multimodal user portrait.

[0028] In an optional implementation, the fusing the user-associated text features, the user-associated image features, and the user-associated behavior features to obtain the multimodal user portrait includes:

[0029] Calling a trained user portrait generation model, fusing the user-associated text features, the user-associated image features, and the user-associated behavior features through the user portrait generation model to obtain the multimodal user portrait;

[0030] The user portrait generation model is obtained by gradually distilling a general large language model.

[0031] In an optional implementation, the user portrait generation model is obtained in the following manner:

[0032] Acquire an unlabeled data set, wherein the unlabeled data set includes multiple groups of user conversation samples and user basic information associated with each group of user conversation samples;

[0033] Inputting the unlabeled data set into the general large language model, and outputting the user portrait label and label explanation reason corresponding to each group of user dialogue samples through the general large language model, wherein the label explanation reason is used to explain the reason for outputting the user portrait label;

[0034] The user portrait generation model is trained by utilizing the multiple groups of user conversation samples, the user basic information associated with each group of user conversation samples, and the user portrait labels corresponding to each group of user conversation samples in combination with the reasons for the label explanations.

[0035] In an optional implementation, extracting topic features associated with the current conversation context includes:

[0036] Performing semantic analysis on the current conversation context to extract conversation key information; the conversation key information includes keywords and / or key phrases;

[0037] Determining the topic category to which the current conversation context belongs according to the conversation key information;

[0038] Extracting hidden user features from the current conversation context; the hidden user features include at least one of the following features of the user: emotional features, interaction patterns, and potential needs;

[0039] A joint feature analysis is performed on the topic category and the hidden user feature to obtain a topic feature associated with the current conversation context.

[0040] In an optional implementation, extracting the conversation intention associated with the current conversation context includes:

[0041] Calling a trained intent classification model, and extracting the conversation intent of the user in the current conversation context through the intent classification model;

[0042] Screening the conversation intentions by importance to obtain key conversation intentions;

[0043] The key conversation intention is determined as the conversation intention associated with the current conversation context.

[0044] In an optional implementation, extracting the conversation intention associated with the current conversation context further includes:

[0045] Performing context coherence analysis on the current conversation context to predict subsequent conversation content;

[0046] extracting the predicted conversation intention of the user in the subsequent conversation content by using the intention classification model;

[0047] An intention evolution sequence formed by arranging the key dialogue intention and the predicted dialogue intention is determined as the dialogue intention associated with the current dialogue context.

[0048] In an optional implementation, extracting the dialog process associated with the current dialog context includes:

[0049] Calling a trained process identification model, parsing the current conversation context through the process identification model to capture hidden conversation features of the current conversation context, wherein the hidden conversation features include at least one of the following features of conversation participants: semantic features, interaction patterns, emotional features, and language features;

[0050] The process identification model is guided by the input guidance information to match the hidden dialogue feature with the process feature corresponding to each preset standard dialogue process, so as to identify the dialogue process associated with the current dialogue context.

[0051] In an optional implementation, outputting a recommended topic for the user through the initial topic recommendation model includes:

[0052] The initial topic recommendation model outputs a recommended topic for the user using a zero-sample learning method based on the multimodal user portrait, the current conversation context and the conversation information.

[0053] In an optional embodiment, the method further includes:

[0054] After outputting the recommended topic for each target user, collecting first interactive feedback data of each target user for the recommended topic; the first interactive feedback data includes at least one of the following: dialogue response behavior, emotional dynamics, and topic preference adjustment data;

[0055] After the data volume of the first interactive feedback data accumulates to a preset data volume threshold, a user preference model is established according to the first interactive feedback data; the user preference model is used to indicate the conversation preference information of each target user, and the conversation preference information includes at least one of the following: the interest of each target user in different topic categories and the conversation style preference of each target user;

[0056] The initial topic recommendation model is fine-tuned based on the conversation preference information of each target user combined with the accumulated conversation data of each target user to obtain an updated topic recommendation model.

[0057] In an optional implementation, after collecting the first interactive feedback data of each target user for the recommended topic, the method further includes:

[0058] evaluating the recommendation effect of the recommended topic according to the first interactive feedback data;

[0059] When the recommendation effect does not meet the preset effect requirement, the recommended topic is updated.

[0060] In an optional implementation, after establishing the user preference model according to the first interactive feedback data, the method further includes:

[0061] The first interactive feedback data is updated regularly, and the conversation preference information of each target user is updated according to the updated first interactive feedback data, so as to update the user preference model.

[0062] In an optional implementation, the recommended topics output by the updated topic recommendation model include a preset proportion of personalized recommended topics generated based on zero-sample learning; the personalized recommended topics have a preset association relationship with the interest points and / or current hot topics of each target user;

[0063] The method further comprises:

[0064] Acquire second interactive feedback data of each target user for the personalized recommended topic;

[0065] evaluating the recommendation effect of the personalized recommended topic according to the second interactive feedback data;

[0066] When the recommendation effect meets the preset effect requirement, the conversation preference information of each target user is updated according to the personalized recommendation topic to update the user preference model.

[0067] According to a second aspect of the present disclosure, there is provided a topic recommendation device, comprising:

[0068] A user portrait determination module, used to determine a multimodal user portrait based on text information associated with the user, image information associated with the user, and behavior information associated with the user;

[0069] A conversation information extraction module, used to obtain the current conversation context of the user and extract the conversation information associated with the current conversation context; the conversation information includes: topic features, conversation intention and conversation process;

[0070] The topic recommendation module is used to input the multimodal user portrait, the current conversation context and the conversation information into an initial topic recommendation model, and output a recommended topic for the user through the initial topic recommendation model.

[0071] According to a third aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the topic recommendation method described in the first aspect is implemented.

[0072] According to a fourth aspect of an embodiment of the present disclosure, there is provided an electronic device, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the topic recommendation method described in the first aspect above by executing the executable instructions.

[0073] According to the topic recommendation method, topic recommendation device, computer-readable storage medium and electronic device of the embodiment of the present disclosure, on the one hand, a multimodal user portrait is determined based on the text information associated with the user, the image information associated with the user and the behavior information associated with the user, so that the user situation can be fully understood from multiple dimensions, thereby portraying a more comprehensive and detailed multimodal user portrait, and providing a basis for subsequent topic recommendations that accurately meet the needs of users; further, by obtaining the current conversation context of the user and extracting the topic features, conversation intentions and conversation progress associated with the current conversation context, the conversation features can reflect the topics involved in the current conversation and their evolution process, the conversation intentions can reflect the user's conversation purpose, and the conversation progress can reflect the topic involved in the current conversation and its evolution process. The process can reflect the current conversation timing, thereby being able to extract more comprehensive conversation information from multiple dimensions, further providing a basis for subsequent topic recommendations that accurately meet user needs; on the other hand, by inputting multimodal user portraits, current conversation contexts and conversation information into the initial topic recommendation model, and outputting recommended topics for users through the initial topic recommendation model, it can ensure that the recommended topics are closely related to the current discussion, avoid irrelevant or repeated information, and provide users with the most demand-oriented topic content, improve the effectiveness and pertinence of the conversation, and ensure that subsequent recommended content is put forward at the right time to avoid interrupting the user's train of thought, thereby making the recommended topics highly in line with user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0075] Figure 1 A schematic diagram showing the flow of a topic recommendation method in an embodiment of the present disclosure is shown;

[0076] Figure 2 A schematic diagram showing a flow chart of how to determine a multimodal user portrait according to text information associated with a user, image information associated with a user, and behavior information associated with a user in an embodiment of the present disclosure;

[0077] Figure 3 A schematic diagram showing a process of how to train the above-mentioned user portrait generation model in an embodiment of the present disclosure;

[0078] Figure 4 An overall schematic diagram showing how to obtain a user portrait generation model in an embodiment of the present disclosure;

[0079] Figure 5 A schematic diagram showing the overall process of generating a multimodal user portrait in an embodiment of the present disclosure;

[0080] Figure 6 A schematic diagram showing a flow chart of how to extract topic features associated with the current conversation context in an embodiment of the present disclosure;

[0081] Figure 7 A schematic diagram showing a first process of extracting a conversation intention associated with a current conversation context in an embodiment of the present disclosure;

[0082] Figure 8 A schematic diagram showing a second flow chart of how to extract a conversation intention associated with a current conversation context in an embodiment of the present disclosure;

[0083] Fig. 9 A schematic diagram showing a flow chart of how to extract a conversation process associated with a current conversation context in an embodiment of the present disclosure;

[0084] Fig.10 A schematic diagram showing a flow chart of how to update the initial topic recommendation model in an embodiment of the present disclosure;

[0085] Fig.11 A flow chart showing how to update the user preference model according to the interactive feedback data of each target user for the personalized recommendation topic in an embodiment of the present disclosure;

[0086] Fig.12 The overall process diagram of the topic recommendation method in the embodiment of the present disclosure is shown;

[0087] Fig.13 A schematic diagram of a topic recommendation device in an embodiment of the present disclosure is shown;

[0088] Fig.14 A structural diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0089] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0090] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0091] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0092] According to an embodiment of the present disclosure, a topic recommendation method, a topic recommendation device, a computer-readable storage medium, and an electronic device are provided.

[0093] In this document, any number of elements in the drawings is used for illustration rather than limitation, and any naming is used only for distinction and does not have any limiting meaning.

[0094] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0096] The inventors have discovered that existing topic recommendation schemes are not accurate enough and cannot meet the actual needs of users, thus affecting the user's interactive experience.

[0097] In view of the above, the basic idea of ​​the present disclosure is to provide a topic recommendation method, a topic recommendation device, a computer-readable storage medium and an electronic device. On the one hand, a multimodal user portrait is determined based on user-associated text information, user-associated image information and user-associated behavior information, so as to fully understand the user situation from multiple dimensions, thereby portraying a more comprehensive and detailed multimodal user portrait, and providing a basis for subsequent topic recommendations that accurately meet user needs; further, by obtaining the user's current conversation context and extracting the topic features, conversation intentions and conversation progress associated with the current conversation context, the conversation features can reflect the topics involved in the current conversation and their evolution process, and the conversation intentions can reflect the user's conversation Purpose: The conversation process can reflect the current conversation timing, so that more comprehensive conversation information can be extracted from multiple dimensions, further providing a basis for subsequent topic recommendations that accurately meet user needs; on the other hand, by inputting multimodal user portraits, current conversation contexts and conversation information into the initial topic recommendation model, the initial topic recommendation model outputs recommended topics for users, which can ensure that the recommended topics are closely related to the current discussion, avoid irrelevant or repeated information, and provide users with the most demand-oriented topic content, improve the effectiveness and pertinence of the conversation, and ensure that subsequent recommended content is put forward at the right time to avoid interrupting the user's train of thought, thereby making the recommended topics highly in line with user needs.

[0098] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0099] Application Scenario Overview

[0100] It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principle of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0101] The embodiments of the present disclosure support topic recommendation in complex real-person social scenarios. For example, in a social APP (Application), text data (user profile, interest tags, historical conversation records), image data (user avatar, shared pictures) and user behavior data (likes, follows, etc.) can be integrated to construct a multimodal user portrait. After that, the user's current conversation context can be obtained, and conversation information including topic features, conversation intentions and conversation progress can be extracted from the current conversation context. The multimodal user portrait, current conversation context and conversation information are input into an initial topic recommendation model, and the initial topic recommendation model outputs recommended topics that can accurately meet user needs.

[0102] Exemplary Methods

[0103] An exemplary embodiment of the present disclosure first provides a topic recommendation method. Figure 1 The flowchart of the topic recommendation method in the embodiment of the present disclosure is shown, which may include the following steps S110 to S130:

[0104] In step S110, a multimodal user portrait is determined based on the text information associated with the user, the image information associated with the user and the behavior information associated with the user.

[0105] In this step, a multimodal user portrait can be determined based on the text information associated with the user, the image information associated with the user, and the behavioral information associated with the user. Multimodal user portrait is a technology that integrates multiple types of data to construct a detailed description of the user. Multimodal user portrait provides a more comprehensive and richer user description by integrating data from different sources and different types. These data can include text, images, videos, audio, sensor data, etc.

[0106] Exemplarily, the text information associated with a user may include at least one of the following: basic information of the user, user tags, historical user conversation data, search history, and social media posts. The basic information of the user may be a personal profile of the user, such as name, gender, occupation, location, education background, professional skills, hobbies, etc.; user tags may be user interest tags (such as favorite music tags, movie tags, sports tags, travel tags, food tags, religious belief tags, etc.), age group tags, personality tags (such as extrovert, introvert, optimistic, pessimistic, etc.), health tags, cultural preference tags, etc.; historical user conversation records may be all conversation contents of the user in the historical period.

[0107] The image information associated with the user may be the user's avatar, photos and visual content uploaded or shared by the user, the time and frequency of the user's viewing, the time and frequency of the recording, etc.; the behavior information associated with the user may be the user's likes, comments, follows and other behaviors, which can be set according to the actual situation, and the present disclosure does not make any special limitations on this.

[0108] refer to Figure 2 , Figure 2 A flow chart showing how to determine a multimodal user portrait according to text information associated with a user, image information associated with a user and behavior information associated with a user in an embodiment of the present disclosure includes steps S201 to S204:

[0109] In step S201, text information associated with a user is obtained, and feature extraction is performed on the text information to obtain user-associated text features.

[0110] In this step, the text information associated with the user may be obtained, and the text information may be feature extracted to obtain the user-associated text features. Exemplarily, the above text information may be deeply analyzed using a general large language model (such as GPT or LLAMA series, Qwen series) to extract the user-associated text features from the above text information. The user-associated text features may be at least one of the following information used to describe the user: user emotional features, user communication style, and user interest features.

[0111] Among them, user emotional characteristics can be user emotional tendencies (for example: positive or negative emotions, active or passive emotions, etc.); user communication style refers to the specific ways and preferences displayed by users in the communication process, including how they express themselves, how they receive information, how they interact with others, etc. For example, user communication style can be direct, indirect, active, passive, organized, unorganized, etc., which can be set according to actual conditions, and the present disclosure does not make special restrictions on this; user interest characteristics can be user's points of interest, such as: technology trends, beauty styling, travel, photography, etc., which can be set according to actual conditions, and the present disclosure does not make special restrictions on this.

[0112] Exemplarily, a text description may be generated based on the extracted user-related text features, for example: "User A prefers to discuss technology trends, and his communication style is rational and organized."

[0113] In step S202, the image information associated with the user is obtained, and feature extraction is performed on the image information to obtain the user-associated image features.

[0114] In this step, the image information associated with the user may be obtained, and the image information may be subjected to feature extraction to obtain the user-associated image features. Exemplarily, a multimodal model (such as CLIP, GPT-4-vision) may be used to analyze the image information associated with the user and extract the user-associated image features. Exemplarily, the user-associated image features may be the style and emotional expression of the image, for example: "The style of the avatar of user A tends to be simple, expressing positive emotions."

[0115] Exemplarily, the user-associated image features may also be converted into a unified text description to facilitate integration with the text description obtained in the above step S201.

[0116] In step S203, the behavior information associated with the user is obtained, and the behavior information is statistically analyzed to obtain the user-associated behavior characteristics.

[0117] In this step, the user-related behavior information can be obtained, and then the behavior information can be statistically analyzed to obtain the user-related behavior characteristics. For example, the user's behavior pattern and content preference can be identified from the user-related behavior information using a statistical analysis method, for example: "User A prefers technology content and interacts frequently."

[0118] Exemplarily, the user-related behavior features may also be converted into a unified text description to form a consistent text format with the aforementioned text and image features, so as to facilitate the subsequent fusion of the three.

[0119] In step S204, the user-associated text features, the user-associated image features, and the user-associated behavior features are fused to obtain a multimodal user portrait.

[0120] In this step, in an optional implementation, the general large language model can be directly called to fuse the user-related text features, user-related image features, and user-related behavior features obtained in the above steps S201 to S203 to obtain a multimodal user portrait. For example, the obtained multimodal user portrait can be: "User A is a young professional with rigorous logic, preference for technology discussions, and a simple aesthetic style."

[0121] In another optional implementation, although the general large language model has a high understanding ability, its huge number of parameters and high computational overhead make it difficult to be efficiently applied in large-scale user information extraction. Therefore, the present disclosure requires a small model with better performance (which can be called a user portrait generation model) to quickly support the generation of multimodal user portraits for a large number of users.

[0122] The two commonly used small model training methods in the industry are fine-tuning and distillation, but these two methods have their own limitations. Specifically, the fine-tuning method requires a large amount of high-quality manually labeled data, which is costly and difficult to obtain continuously. Although the distillation method can save labeling costs by training through the output of a large model, it still requires a large amount of unlabeled data, and there are also bottlenecks in data acquisition.

[0123] To solve the above problems, the present disclosure adopts a step-by-step distillation mechanism. Compared with traditional methods, step-by-step distillation can utilize the reasoning ability of large models and efficiently train small models with less data.

[0124] Exemplary, reference Figure 3 , Figure 3 A schematic diagram of a process of how to train the user portrait generation model in the embodiment of the present disclosure is shown, including steps S301 to S303:

[0125] In step S301, an unlabeled data set is obtained.

[0126] In this step, an unlabeled data set may be obtained, and the unlabeled data includes multiple groups of user dialogue samples and user basic information associated with each group of user dialogue samples. Exemplarily, the user dialogue samples may be historical dialogue records between a user and other users, or historical dialogue records between a user and other users in an online forum or social media interactive platform, or historical dialogue records between a user and a customer service robot, intelligent assistant, etc., which may be set according to actual conditions, and the present disclosure does not specifically limit this. The user basic information associated with each group of user dialogue samples may include the name, gender, age, occupation, points of interest, etc. of the main perspective user in the user dialogue sample (for example: assuming that three groups of user dialogue samples are obtained for user A, the first group of user dialogue samples are dialogue samples between user A and user B, the second group of user dialogue samples are dialogue samples between user A and user C, and the third group of user dialogue samples are dialogue samples between user A and user D, then user A is the main perspective user).

[0127] In step S302, the unlabeled data set is input into a general large language model, and the general large language model outputs the user portrait label and label explanation reason corresponding to each group of user dialogue samples.

[0128] In this step, the above-mentioned unlabeled data set can be input into a general large language model, and the general large language model can output the user portrait label and label explanation reason corresponding to each group of user dialogue samples. The label explanation reason is used to explain the reason for outputting the user portrait label.

[0129] Among them, the user portrait label is the portrait information output by the general large language model based on each group of user conversation samples, and the "label explanation reason" refers to providing a detailed description for each user label, explaining why the user is labeled and the specific meaning of the label. This approach helps to ensure the accuracy and consistency of the labels, and provides a clear understanding and guidance for subsequent models that use these labels, eliminating the analysis process of subsequent models and improving model training efficiency.

[0130] In step S303, a user portrait generation model is trained using multiple groups of user conversation samples, user portrait labels corresponding to each group of user conversation samples, and reasons for the label explanations.

[0131] In this step, the above-mentioned multiple groups of user dialogue samples, the user portrait labels corresponding to each group of user dialogue samples and their label explanation reasons can be used as new training samples to train and obtain a user portrait generation model.

[0132] After the user portrait generation model is trained, the present disclosure can call the trained user portrait generation model, and fuse user-associated text features, user-associated image features, and user-associated behavior features through the user portrait generation model to obtain a multimodal user portrait.

[0133] Therefore, the present disclosure can quickly distill a user portrait generation model with better performance and more suitable for multimodal user portrait generation scenarios that support a large number of users from a general large model, thereby avoiding the problem of high computational overhead caused by using a general large model to generate multimodal user portraits.

[0134] refer to Figure 4 , Figure 4 The overall schematic diagram of how to obtain the user portrait generation model in the embodiment of the present disclosure is shown as follows: Figure 4 As shown:

[0135] First, we can generate portrait labels and label explanations based on the unlabeled dataset (including multiple groups of user dialogue samples and user basic information associated with each group of user dialogue samples, where the user basic information includes user interests) using the Prompt-guided general large language model (LLM).

[0136] Afterwards, multiple groups of user conversation samples, basic user information associated with each group of user conversation samples, user portrait labels corresponding to each group of user conversation samples, and their label explanation reasons can be used as new training samples to fine-tune the user portrait generation model SLM.

[0137] refer to Figure 5 , Figure 5 The overall process diagram of how to generate a multimodal user portrait in an embodiment of the present disclosure is shown, including steps S501 to S506:

[0138] In step S501, multimodal data is collected, wherein the multimodal data includes user-associated text information, user-associated image information, and user-associated behavior information;

[0139] In step S502, feature extraction is performed on the user-associated text information through a general large prediction model to obtain user-associated text features;

[0140] In step S503, feature extraction is performed on the user-associated image information through the multimodal large model to obtain user-associated image features;

[0141] In step S504, statistical analysis is performed on the user-associated behavior information to obtain user-associated behavior characteristics;

[0142] In step S505, the description of the association of each dimension of the user output in text form is obtained;

[0143] In step S506, the user portrait generation model is guided by the input guidance information to fuse the user-associated text features, the user-associated image features and the user-associated behavior features to obtain a multimodal user portrait.

[0144] After obtaining the multimodal user portrait, the process may proceed to step S120 to obtain the user's current conversation context and extract conversation information associated with the current conversation context.

[0145] In this step, the user's current conversation context can be obtained, and the topic features, conversation intentions, and conversation progress associated with the current conversation context can be extracted.

[0146] Among them, the current conversation context can be the current multi-round conversation context between the user and a specified contact. The topic feature refers to the attributes of the subject or content involved in the conversation, which is used to describe the main discussion points of the current conversation. The conversation intention is the purpose or demand expressed by the user in the conversation, reflecting the goal that the user hopes to achieve through the conversation. The conversation process refers to the development stage or step of the conversation, which reflects the state changes and progress of the conversation.

[0147] refer to Figure 6 , Figure 6 A flow chart showing how to extract topic features associated with the current conversation context in an embodiment of the present disclosure includes steps S601 to S604:

[0148] In step S601, semantic analysis is performed on the current conversation context to extract key conversation information.

[0149] In this step, illustratively, a general large language model (LLM) may be used to perform semantic analysis on the current conversation context to extract key conversation information, wherein the key conversation information may include keywords and / or key phrases. illustratively, the keywords and / or key phrases may be high-frequency words or phrases that appear repeatedly in the conversation, and may be set according to actual conditions, and the present disclosure does not make any special limitation on this.

[0150] In step S602, the topic category to which the current conversation context belongs is determined according to the conversation key information.

[0151] In this step, the above-mentioned conversation key information can be mapped to a preset topic category to determine the topic category to which the current conversation context belongs. Exemplarily, the topic category can be interests, hobbies, emotional states, learning goals, etc.

[0152] In step S603, hidden user features are extracted from the current conversation context.

[0153] In this step, hidden user features can be extracted from the current conversation context, and the hidden user features include at least one of the following features of the user: emotional features, interaction patterns, and potential needs. Among them, emotional features refer to the emotional state or emotional tendency shown by the user in the conversation, such as happiness, anger, anxiety, sadness, etc.; interaction patterns refer to the behavioral patterns and communication methods of the conversation between the user and the system, including inquiries, suggestions, question frequency, answer length, conversation rhythm, etc.; potential needs refer to the actual needs or expectations of the user that are not clearly expressed in the conversation but can be inferred through analysis.

[0154] In step S604, a joint feature analysis is performed on the topic category and the hidden user feature to obtain the topic feature associated with the current conversation context.

[0155] In this step, based on the named entity recognition (NER) and sentiment analysis capabilities of LLM, the above topic categories and hidden user features can be jointly analyzed to obtain topic features associated with the current conversation context. Exemplarily, the NER capability of LLM can be used to identify key entities in the topic category, and the sentiment analysis capability of LLM can be used to identify the emotional state in the conversation. Then, the results of NER and sentiment analysis can be combined to perform joint feature analysis to mine topic features associated with the current conversation context.

[0156] Exemplarily, the present disclosure may extract the conversation intention associated with the current conversation context based on the following two implementations.

[0157] In a first optional implementation, reference Figure 7 , Figure 7 A schematic diagram of a process of extracting a conversation intention associated with a current conversation context in a first embodiment of the present disclosure is shown, including steps S701 to S703:

[0158] In step S701, the trained intent classification model is called to extract the user's conversation intent in the current conversation context through the intent classification model.

[0159] In this step, the trained intent classification model can be called to extract the user's conversation intent in the current conversation context through the intent classification model. Exemplarily, the intent classification model can first identify the user's intent from the current conversation context, and then map the identified user's intent to a preset intent category (for example, seeking advice, emotional expression, social interaction, etc.), and output the user's conversation intent in the current conversation context. Given that the current conversation context may contain multiple rounds of conversations or multiple conversation sentences, one or more conversation intents can be extracted. Conversation intent refers to the specific purpose or need expressed by the user in the conversation.

[0160] In step S702, the conversation intentions are screened by importance to obtain key conversation intentions.

[0161] In this step, when multiple conversation intentions are obtained in the above step S701, the importance of the above multiple conversation intentions can be screened based on the weight evaluation mechanism to obtain the key conversation intentions.

[0162] In step S703, the key conversation intention is determined as the conversation intention associated with the current conversation context.

[0163] In this step, the above key conversation intention can be determined as the conversation intention associated with the current conversation context. Based on this implementation, the present disclosure can screen out the key conversation intention of the user in the conversation process, thereby avoiding the interference of some unimportant conversation intentions on the subsequent topic recommendation process, and facilitating the subsequent accurate topic recommendation based on the key conversation intention.

[0164] In a second optional implementation, after obtaining the above key conversation intention, refer to Figure 8 , Figure 8 A schematic diagram of a second process of extracting a conversation intention associated with the current conversation context in an embodiment of the present disclosure is shown, including steps S801 to S803:

[0165] In step S801, a context coherence analysis is performed on the current conversation context to predict the subsequent conversation content.

[0166] In this step, the user's speech changes in the current conversation context can be tracked, and the contextual coherence analysis of the current conversation context can be performed to predict the subsequent conversation content. Contextual coherence analysis refers to evaluating and ensuring the semantic consistency and logical coherence between the various parts of the conversation. It involves understanding the conversation history and predicting and generating appropriate subsequent content or responses based on this historical information.

[0167] In step S802, the intention classification model is used to extract the user's predicted conversation intention in the subsequent conversation content.

[0168] In this step, after predicting the subsequent conversation content, the trained intent classification model can be called, and the intent classification model can be used to extract the user's predicted conversation intention in the subsequent conversation content, so that the user's possible purpose or needs in the subsequent conversation can be predicted in advance.

[0169] In step S803, the intention evolution sequence formed by arranging the key dialogue intentions and the predicted dialogue intentions is determined as the dialogue intention associated with the current dialogue context.

[0170] In this step, the key conversation intention obtained in the above step S703 and the predicted conversation intention obtained in the above step S802 can be arranged in chronological order to obtain an intention evolution sequence, and the intention evolution sequence can be determined as the conversation intention associated with the current conversation context.

[0171] Based on this implementation, the present disclosure can determine the changing trend of the user's intention during the conversation, so as to facilitate the subsequent recommendation of more targeted topics to the user based on the changing trend of the intention.

[0172] refer to Fig. 9 , Fig. 9 A flowchart showing how to extract a conversation process associated with a current conversation context in an embodiment of the present disclosure includes steps S901 to S902:

[0173] In step S901, the trained process identification model is called to parse the current dialog context through the process identification model to capture the hidden dialog features of the current dialog context.

[0174] In this step, the trained process identification model can be called to analyze the current conversation context through the process identification model to capture the hidden conversation features of the current conversation context. The hidden conversation features can include at least one of the following features of the conversation participants: semantic features, interaction patterns, emotional features, and language features.

[0175] Among them, semantic features can refer to the attributes of the meaning or implications expressed by the conversation content, which involves the characteristics of words, sentences or paragraphs at the semantic level; emotional features refer to the emotional state or emotional tendency shown by the user in the conversation, such as happiness, anger, anxiety, sadness, etc.; interaction mode refers to the behavioral pattern and communication method of the conversation between the user and the system, including inquiries, suggestions, question frequency, answer length, conversation rhythm, etc.; language features refer to the characteristics of the conversation content in terms of grammar, syntax, vocabulary usage, etc., and it mainly focuses on the form and structure of the text.

[0176] In step S902, the input guidance information guides the process identification model to match the hidden dialogue features with the process features corresponding to each preset standard dialogue process, so as to identify the dialogue process associated with the current dialogue context.

[0177] In this step, after obtaining the above hidden dialogue features, the hidden dialogue features can be matched with the process features corresponding to each preset standard dialogue process through the input guidance information Prompt to guide the process recognition model to identify the dialogue process associated with the current dialogue context.

[0178] Exemplarily, the preset standard dialogue process may be, for example, "initial greetings", "interest discussion", "emotional exchange", "conversation ending", etc. The above-mentioned guidance information Prompt may be set, for example, to match its hidden dialogue features with each preset standard dialogue process according to the following dialogue content, and output the dialogue process associated with the current dialogue context.

[0179] After obtaining the multimodal user portrait, topic features associated with the current conversation context, conversation intention and conversation process, step S130 can be entered to input the multimodal user portrait, current conversation context and conversation information into the initial topic recommendation model, and the recommended topic for the user is output through the initial topic recommendation model.

[0180] In this step, the above-mentioned multimodal user portrait, current conversation context and conversation information (topic features, conversation intention and conversation process) can be input into the initial topic recommendation model, and the recommended topic for the user can be output through the initial topic recommendation model.

[0181] For example, the initial topic recommendation model can output recommended topics for users based on multimodal user portraits, current conversation context and conversation information using zero-shot learning. Zero-shot learning (ZSL) is a machine learning method that allows the model to predict or classify specific categories without having seen training data for these categories.

[0182] Therefore, the present disclosure can effectively deal with cold start situations, and in the absence of user historical conversation data, it can also output diversified recommended topics that are consistent with user personality labels and cover the user's potential areas of interest.

[0183] It should be noted that, with the long-term use and data accumulation of the initial topic recommendation model, the present disclosure can also update the above initial topic recommendation model according to the accumulated user data to improve the performance of the model. Fig.10 , Fig.10 A schematic diagram showing a flow chart of how to update the initial topic recommendation model in an embodiment of the present disclosure, including steps S1001 to S1003:

[0184] In step S1001, after outputting a recommended topic for each target user, first interactive feedback data of each target user for the recommended topic is collected.

[0185] In this step, after outputting the recommended topic for each target user, the first interactive feedback data of each target user for the recommended topic is collected. Interactive feedback data refers to feedback information generated by the user during the interaction with the recommended topic. Exemplarily, the above-mentioned first interactive feedback data may include at least one of the following: dialogue response behavior, emotional dynamics, and topic preference adjustment data. Among them, the dialogue response behavior can be the specific response method and pattern shown by the user to the recommended topic during the dialogue, including the topic response rate (for example: click-through rate, conversion rate), response speed, response frequency, response content, response form, stay time, etc.; emotional dynamics can be the changes in the user's emotional state during the dialogue, including the type and intensity of emotion and its evolution over time; topic preference adjustment data can be the dialogue or related operations entered by the target user in the dialog box of the recommended topic to adjust the recommended topic.

[0186] In step S1002, after the data volume of the first interactive feedback data reaches a preset data volume threshold, a user preference model is established according to the first interactive feedback data.

[0187] In this step, after the data volume of the first interactive feedback data has accumulated to a preset data volume threshold, a user preference model may be established based on the first interactive feedback data.

[0188] The above-mentioned preset data volume threshold can be set according to actual needs, and the present disclosure does not make any special restrictions on this. The above-mentioned user preference model is used to indicate the conversation preference information of each target user, and the conversation preference information may include at least one of the following: each target user's interest in different topic categories (for example: entertainment, technology, social, etc.) and each target user's conversation style preference (for example: serious, humorous, casual, etc.).

[0189] In step S1003, the initial topic recommendation model is fine-tuned based on the conversation preference information of each target user combined with the accumulated conversation data of each target user to obtain an updated topic recommendation model.

[0190] In this step, the initial topic recommendation model can be fine-tuned based on the conversation preference information of each target user combined with the accumulated conversation data of each target user to obtain an updated topic recommendation model.

[0191] Therefore, the fine-tuned topic recommendation model can output more accurate topic content that meets user preferences in the process of generating recommended topics based on user preferences combined with multimodal user portraits, topic features, conversation intentions, and conversation progress information, making the recommended topics highly consistent with the conversation scenarios.

[0192] In an optional implementation, after collecting the first interactive feedback data of each target user for the recommended topic, the present disclosure may also evaluate the recommendation effect of the recommended topic based on the first interactive feedback data, and update the recommended topic if the recommendation effect does not meet the preset effect requirements.

[0193] Exemplarily, the above-mentioned dialogue response behavior, emotional dynamics and topic preference adjustment data can be evaluated to determine the target user's satisfaction score for the recommended topic, and then use the satisfaction score as the evaluated recommendation effect. For example, when the target user's dialogue response behavior is to click and take the recommended topic within 2 seconds, the indicator score of the dialogue response behavior can be determined to be 8 points, when the emotional dynamics is a positive response, the score of the target user's emotional dynamics can be determined to be 8 points, and when the target user's topic preference adjustment data is an interactive operation that does not require adjustment of the topic, the score of the target user's topic preference adjustment data can be determined to be 9 points. Thus, the above-mentioned scores can be accumulated or weighted to determine the overall satisfaction score (i.e., the recommendation effect of the recommended topic).

[0194] Afterwards, the recommendation effect can be compared with the above-mentioned preset effect requirements (which can be a preset satisfaction score threshold). If the recommendation effect does not meet the preset effect requirements, the recommended topic can be updated in combination with information such as the current conversation context and the current conversation process. It can be timely perceived whether the recommended topic accurately reaches the user, so as to make timely adjustments to the recommended topic so that the recommendation can match the conversation process and ensure the continuity of the conversation and the smoothness of the interaction.

[0195] In an optional implementation, after establishing the user preference model according to the first interactive feedback data, the present disclosure may also periodically analyze the first interactive feedback data of the target user on the recommended topic, and timely update the first interactive feedback data, and then update the conversation preference information of each target user according to the updated first interactive feedback data to update the user preference model. Thus, the present disclosure can realize that the user preference model changes with the dynamic changes of the user's conversation preference, thereby further improving the accuracy of the subsequent recommended topics.

[0196] In an optional implementation, the recommended topics output by the updated topic recommendation model of the present disclosure may include a preset proportion (for example, 20%-30%, which can be set according to actual conditions and is not specifically limited by the present disclosure) of personalized recommended topics generated based on zero-sample learning, and the personalized recommended topics have a preset correlation with each target user's interest points and / or current hot topics.

[0197] Therefore, the present disclosure can also update the user preference model according to the interactive feedback data of each target user for the personalized recommendation topic. For details, please refer to Fig.11 , Fig.11 A flow chart showing how to update the user preference model according to the interactive feedback data of each target user for the personalized recommendation topic in the embodiment of the present disclosure includes steps S1101 to S1103:

[0198] In step S1101, second interactive feedback data of each target user for the personalized recommendation topic is obtained.

[0199] In this step, after the recommended topics are output to the target users through the updated topic recommendation model, given that the recommended topics include some personalized recommended topics, the second interactive feedback data of each target user for the personalized recommended topics can be obtained. Exemplarily, referring to the relevant explanation of the above-mentioned first interactive feedback data, the second interactive feedback data may also include at least one of the following: dialogue response behavior, emotional dynamics, and topic preference adjustment data.

[0200] In step S1102, the recommendation effect of the personalized recommended topic is evaluated according to the second interactive feedback data.

[0201] In this step, the recommendation effect of the personalized recommendation topic can be evaluated based on the second interactive feedback data. Exemplarily, referring to the relevant explanations of the above steps, the satisfaction score of the target user for the personalized recommendation topic can be determined based on the dialogue response behavior, emotional dynamics and topic preference adjustment data, and the satisfaction score can be regarded as the recommendation effect.

[0202] In step S1103, when the recommendation effect meets the preset effect requirement, the conversation preference information of each target user is updated according to the personalized recommendation topic to update the user preference model.

[0203] In this step, when the evaluated recommendation effect meets the preset effect requirements (for example, the satisfaction score is greater than or equal to the preset satisfaction score threshold), the conversation preference information of each target user can be updated according to the personalized recommendation topic to update the above-mentioned user preference model. Thus, the present disclosure can gradually form a closed loop of "exploration-evaluation-optimization".

[0204] refer to Fig.12 , Fig.12 The overall flow chart of the topic recommendation method in the embodiment of the present disclosure is shown, including steps S1201 to S1208:

[0205] In step S1201, the user's current conversation context is obtained;

[0206] In step S1202, topic features associated with the current conversation context are extracted;

[0207] In step S1203, extracting the conversation intention associated with the current conversation context;

[0208] In step S1204, the conversation process associated with the current conversation context is extracted;

[0209] In step S1205, the current conversation context, topic features, conversation intention, conversation process and multimodal user portrait are input into the initial topic recommendation model;

[0210] In step S1206, obtaining recommended topics;

[0211] In step S1207, user interaction feedback data is collected;

[0212] In step S1208, the initial topic recommendation model is fine-tuned to obtain an updated topic recommendation model.

[0213] Based on the above technical solutions, the present disclosure has at least the following technical effects:

[0214] First, a complete solution for social topic recommendation based on a large model is proposed. By supporting the efficient generation of massive multimodal portraits, real-time context analysis and intent recognition combined with personalized generative topic recommendation, it can effectively address the dynamically changing user needs in real-life social scenarios.

[0215] Second, using multimodal data (text, image, and behavior data) to build user portraits can facilitate in-depth mining of users' emotional characteristics, personality tendencies, aesthetic preferences, and other personalized information based on user portraits;

[0216] Third, a two-stage topic generation scheme is proposed. In the initial stage, a zero-shot method is used to quickly generate diverse initial topics. Then, as data accumulates, the accuracy of topic generation is improved through model fine-tuning. At the same time, some zero-shot generated content is retained, and personalized recommendation topics are introduced to avoid topic monotony, thereby ensuring the personalization, richness and exploratory nature of the recommended content.

[0217] Exemplary Devices

[0218] After introducing the topic recommendation method according to the exemplary embodiment of the present disclosure, Fig.13 A topic recommendation device according to an exemplary embodiment of the present disclosure will be described.

[0219] Fig.13 FIG. 4 shows a schematic diagram of the structure of a topic recommendation device in an exemplary embodiment of the present disclosure; Fig.13As shown, the topic recommendation device 1300 may include a user portrait determination module 1310, a conversation information extraction module 1320 and a topic recommendation module 1330. Among them:

[0220] A user portrait determination module 1310 is used to determine a multimodal user portrait based on text information associated with the user, image information associated with the user, and behavior information associated with the user;

[0221] The conversation information extraction module 1320 is used to obtain the current conversation context of the user and extract the conversation information associated with the current conversation context; the conversation information includes: topic features, conversation intention and conversation process;

[0222] The topic recommendation module 1330 is used to input the multimodal user portrait, the current conversation context and the conversation information into an initial topic recommendation model, and output a recommended topic for the user through the initial topic recommendation model.

[0223] In an optional implementation, the user portrait determination module 1310 determines a multimodal user portrait according to the text information associated with the user, the image information associated with the user, and the behavior information associated with the user, including:

[0224] Acquire text information associated with the user, and perform feature extraction on the text information to obtain the user-associated text feature; the text information includes at least one of the following: basic information of the user, user tags, and user historical conversation data; the user-associated text feature is used to describe at least one piece of information of the user: user emotional features, user communication style, and user interest features;

[0225] Acquire image information associated with the user, and perform feature extraction on the image information to obtain features of the user associated image;

[0226] Acquire the behavior information associated with the user, and perform statistical analysis on the behavior information to obtain the behavior characteristics associated with the user;

[0227] The user-associated text features, the user-associated image features, and the user-associated behavior features are fused to obtain the multimodal user portrait.

[0228] In an optional implementation, the user portrait determination module 1310 fuses the user-associated text features, the user-associated image features, and the user-associated behavior features to obtain the multimodal user portrait, including:

[0229] Calling a trained user portrait generation model, fusing the user-associated text features, the user-associated image features, and the user-associated behavior features through the user portrait generation model to obtain the multimodal user portrait;

[0230] The user portrait generation model is obtained by gradually distilling a general large language model.

[0231] In an optional implementation, the user portrait generation model is obtained in the following manner:

[0232] Acquire an unlabeled data set, wherein the unlabeled data set includes multiple groups of user conversation samples and user basic information associated with each group of user conversation samples;

[0233] Inputting the unlabeled data set into the general large language model, and outputting the user portrait label and label explanation reason corresponding to each group of user dialogue samples through the general large language model, wherein the label explanation reason is used to explain the reason for outputting the user portrait label;

[0234] The user portrait generation model is trained by utilizing the multiple groups of user conversation samples, the user basic information associated with each group of user conversation samples, and the user portrait labels corresponding to each group of user conversation samples in combination with the reasons for the label explanations.

[0235] In an optional implementation, the conversation information extraction module 1320 extracts topic features associated with the current conversation context, including:

[0236] Performing semantic analysis on the current conversation context to extract conversation key information; the conversation key information includes keywords and / or key phrases;

[0237] Determining the topic category to which the current conversation context belongs according to the conversation key information;

[0238] Extracting hidden user features from the current conversation context; the hidden user features include at least one of the following features of the user: emotional features, interaction patterns, and potential needs;

[0239] A joint feature analysis is performed on the topic category and the hidden user feature to obtain a topic feature associated with the current conversation context.

[0240] In an optional implementation, the dialogue information extraction module 1320 extracts the dialogue intention associated with the current dialogue context, including:

[0241] Calling a trained intent classification model, and extracting the conversation intent of the user in the current conversation context through the intent classification model;

[0242] Screening the conversation intentions by importance to obtain key conversation intentions;

[0243] The key conversation intention is determined as the conversation intention associated with the current conversation context.

[0244] In an optional implementation, the dialog information extraction module 1320 extracts the dialog intention associated with the current dialog context, further comprising:

[0245] Performing context coherence analysis on the current conversation context to predict subsequent conversation content;

[0246] extracting the predicted conversation intention of the user in the subsequent conversation content by using the intention classification model;

[0247] An intention evolution sequence formed by arranging the key dialogue intention and the predicted dialogue intention is determined as the dialogue intention associated with the current dialogue context.

[0248] In an optional implementation, the conversation information extraction module 1320 extracts the conversation process associated with the current conversation context, including:

[0249] Calling a trained process identification model, parsing the current conversation context through the process identification model to capture hidden conversation features of the current conversation context, wherein the hidden conversation features include at least one of the following features of conversation participants: semantic features, interaction patterns, emotional features, and language features;

[0250] The process identification model is guided by the input guidance information to match the hidden dialogue feature with the process feature corresponding to each preset standard dialogue process, so as to identify the dialogue process associated with the current dialogue context.

[0251] In an optional implementation, the topic recommendation module 1330 outputs a recommended topic for the user through the initial topic recommendation model, including:

[0252] The initial topic recommendation model outputs a recommended topic for the user using a zero-sample learning method based on the multimodal user portrait, the current conversation context and the conversation information.

[0253] In an optional implementation, the topic recommendation module 1330 is configured to:

[0254] After outputting the recommended topic for each target user, collecting first interactive feedback data of each target user for the recommended topic; the first interactive feedback data includes at least one of the following: dialogue response behavior, emotional dynamics, and topic preference adjustment data;

[0255] After the data volume of the first interactive feedback data accumulates to a preset data volume threshold, a user preference model is established according to the first interactive feedback data; the user preference model is used to indicate the conversation preference information of each target user, and the conversation preference information includes at least one of the following: the interest of each target user in different topic categories and the conversation style preference of each target user;

[0256] The initial topic recommendation model is fine-tuned based on the conversation preference information of each target user combined with the accumulated conversation data of each target user to obtain an updated topic recommendation model.

[0257] In an optional implementation, after collecting the first interactive feedback data of each target user for the recommended topic, the topic recommendation module 1330 is configured to:

[0258] evaluating the recommendation effect of the recommended topic according to the first interactive feedback data;

[0259] When the recommendation effect does not meet the preset effect requirement, the recommended topic is updated.

[0260] In an optional implementation, after establishing the user preference model according to the first interactive feedback data, the topic recommendation module 1330 is configured to:

[0261] The first interactive feedback data is updated regularly, and the conversation preference information of each target user is updated according to the updated first interactive feedback data, so as to update the user preference model.

[0262] In an optional implementation, the recommended topics output by the updated topic recommendation model include a preset proportion of personalized recommended topics generated based on zero-sample learning; the personalized recommended topics have a preset association relationship with the interest points and / or current hot topics of each target user;

[0263] The topic recommendation module 1330 is configured to:

[0264] Acquire second interactive feedback data of each target user for the personalized recommended topic;

[0265] evaluating the recommendation effect of the personalized recommended topic according to the second interactive feedback data;

[0266] When the recommendation effect meets the preset effect requirement, the conversation preference information of each target user is updated according to the personalized recommendation topic to update the user preference model.

[0267] In addition, other specific details of the embodiments of the present disclosure have been described in detail in the inventive implementation of the above method and will not be repeated here.

[0268] Exemplary Storage Media

[0269] The storage medium according to the exemplary embodiment of the present disclosure is described below.

[0270] In this exemplary embodiment, the above method can be implemented by a program product, such as a portable compact disk read-only memory (CD-ROM) and including program code, and can be run on a device, such as a personal computer. However, the program product of the present disclosure is not limited thereto, and in this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus, or a device.

[0271] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0272] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0273] The program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RE, etc., or any suitable combination of the foregoing.

[0274] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (FAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0275] Exemplary Electronic Devices

[0276] refer to Fig.14 An electronic device according to an exemplary embodiment of the present disclosure is described.

[0277] Fig.14 The electronic device 1400 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0278] like Fig.14 As shown, the electronic device 1400 is in the form of a general computing device. The components of the electronic device 1400 may include but are not limited to: at least one processor 1410, at least one memory 1420, a bus 1430 connecting different system components (including the memory 1420 and the processor 1410), and a display unit 1440.

[0279] The memory stores program codes, which can be executed by the processor 1410, so that the processor 1410 executes the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processor 1410 can execute the following steps: Figure 1 The method steps shown, etc.

[0280] The memory 1420 may include a volatile memory, such as a random access memory (RAM) 1421 and / or a cache memory 1422 , and may further include a read-only memory (ROM) 1423 .

[0281] Memory 1420 may also include a program / utility 1424 having a set (at least one) of program modules 1425, such program modules 1425 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0282] The bus 1430 may include a data bus, an address bus, and a control bus.

[0283] The electronic device 1400 may also communicate with one or more external devices 1500 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), and such communication may be performed via an input / output (I / O) interface 1450. The electronic device 1400 also includes a display unit 1440, which is connected to the input / output (I / O) interface 1450 for display. In addition, the electronic device 1400 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 1460. As shown, the network adapter 1460 communicates with other modules of the electronic device 1400 via a bus 1430. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 1400, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0284] It should be noted that, although several modules or submodules of the device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0285] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0286] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A topic recommendation method, characterized in that: include: Determine a multimodal user portrait based on the text information associated with the user, the image information associated with the user, and the behavior information associated with the user; Acquire the current conversation context of the user, and extract conversation information associated with the current conversation context; The conversation information includes: topic features, conversation intention and conversation process; The multimodal user portrait, the current conversation context and the conversation information are input into an initial topic recommendation model, and a recommended topic for the user is output through the initial topic recommendation model.

2. The method according to claim 1, characterized in that The determining of the multimodal user portrait according to the text information associated with the user, the image information associated with the user and the behavior information associated with the user includes: Acquire text information associated with the user, and perform feature extraction on the text information to obtain the user-associated text feature; the text information includes at least one of the following: basic information of the user, user tags, and user historical conversation data; the user-associated text feature is used to describe at least one piece of information of the user: user emotional features, user communication style, and user interest features; Acquire image information associated with the user, and perform feature extraction on the image information to obtain features of the user associated image; Acquire the behavior information associated with the user, and perform statistical analysis on the behavior information to obtain the behavior characteristics associated with the user; The user-associated text features, the user-associated image features, and the user-associated behavior features are fused to obtain the multimodal user portrait.

3. The method according to claim 1, characterized in that The fusing the user-associated text features, the user-associated image features, and the user-associated behavior features to obtain the multimodal user portrait includes: Calling a trained user portrait generation model, fusing the user-associated text features, the user-associated image features, and the user-associated behavior features through the user portrait generation model to obtain the multimodal user portrait; The user portrait generation model is obtained by gradually distilling a general large language model.

4. The method according to claim 3, characterized in that The user portrait generation model is obtained in the following way: Acquire an unlabeled data set, wherein the unlabeled data set includes multiple groups of user conversation samples and user basic information associated with each group of user conversation samples; Inputting the unlabeled data set into the general large language model, and outputting the user portrait label and label explanation reason corresponding to each group of user dialogue samples through the general large language model, wherein the label explanation reason is used to explain the reason for outputting the user portrait label; The user portrait generation model is trained by utilizing the multiple groups of user conversation samples, the user basic information associated with each group of user conversation samples, and the user portrait labels corresponding to each group of user conversation samples in combination with the reasons for the label explanations.

5. The method according to claim 1, characterized in that Extracting topic features associated with the current conversation context includes: Performing semantic analysis on the current conversation context to extract conversation key information; the conversation key information includes keywords and / or key phrases; Determining the topic category to which the current conversation context belongs according to the conversation key information; Extracting hidden user features from the current conversation context; the hidden user features include at least one of the following features of the user: emotional features, interaction patterns, and potential needs; A joint feature analysis is performed on the topic category and the hidden user feature to obtain a topic feature associated with the current conversation context.

6. The method according to claim 1, characterized in that Extracting the conversation intention associated with the current conversation context, including: Calling a trained intent classification model, and extracting the conversation intent of the user in the current conversation context through the intent classification model; Screening the conversation intentions by importance to obtain key conversation intentions; The key conversation intention is determined as the conversation intention associated with the current conversation context.

7. The method according to claim 6, characterized in that Extracting the conversation intention associated with the current conversation context also includes: Performing context coherence analysis on the current conversation context to predict subsequent conversation content; extracting the predicted conversation intention of the user in the subsequent conversation content by using the intention classification model; An intention evolution sequence formed by arranging the key dialogue intention and the predicted dialogue intention is determined as the dialogue intention associated with the current dialogue context.

8. A topic recommendation device, characterized in that: include: A user portrait determination module, used to determine a multimodal user portrait based on text information associated with the user, image information associated with the user, and behavior information associated with the user; A conversation information extraction module, used to obtain the current conversation context of the user and extract conversation information associated with the current conversation context; The conversation information includes: topic features, conversation intention and conversation process; The topic recommendation module is used to input the multimodal user portrait, the current conversation context and the conversation information into an initial topic recommendation model, and output a recommended topic for the user through the initial topic recommendation model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: processor; as well as A memory, configured to store executable instructions of the processor; The processor is configured to perform the method according to any one of claims 1 to 7 by executing the executable instructions.

Images